| 1 | #include "ColumnVector.h" |
| 2 | |
| 3 | #include <cstring> |
| 4 | #include <cmath> |
| 5 | #include <common/unaligned.h> |
| 6 | #include <Common/Exception.h> |
| 7 | #include <Common/Arena.h> |
| 8 | #include <Common/SipHash.h> |
| 9 | #include <Common/NaNUtils.h> |
| 10 | #include <Common/RadixSort.h> |
| 11 | #include <Common/assert_cast.h> |
| 12 | #include <IO/WriteBuffer.h> |
| 13 | #include <IO/WriteHelpers.h> |
| 14 | #include <Columns/ColumnsCommon.h> |
| 15 | #include <DataStreams/ColumnGathererStream.h> |
| 16 | #include <ext/bit_cast.h> |
| 17 | #include <pdqsort.h> |
| 18 | |
| 19 | #ifdef __SSE2__ |
| 20 | #include <emmintrin.h> |
| 21 | #endif |
| 22 | |
| 23 | namespace DB |
| 24 | { |
| 25 | |
| 26 | namespace ErrorCodes |
| 27 | { |
| 28 | extern const int PARAMETER_OUT_OF_BOUND; |
| 29 | extern const int SIZES_OF_COLUMNS_DOESNT_MATCH; |
| 30 | } |
| 31 | |
| 32 | |
| 33 | template <typename T> |
| 34 | StringRef ColumnVector<T>::serializeValueIntoArena(size_t n, Arena & arena, char const *& begin) const |
| 35 | { |
| 36 | auto pos = arena.allocContinue(sizeof(T), begin); |
| 37 | unalignedStore<T>(pos, data[n]); |
| 38 | return StringRef(pos, sizeof(T)); |
| 39 | } |
| 40 | |
| 41 | template <typename T> |
| 42 | const char * ColumnVector<T>::deserializeAndInsertFromArena(const char * pos) |
| 43 | { |
| 44 | data.push_back(unalignedLoad<T>(pos)); |
| 45 | return pos + sizeof(T); |
| 46 | } |
| 47 | |
| 48 | template <typename T> |
| 49 | void ColumnVector<T>::updateHashWithValue(size_t n, SipHash & hash) const |
| 50 | { |
| 51 | hash.update(data[n]); |
| 52 | } |
| 53 | |
| 54 | template <typename T> |
| 55 | struct ColumnVector<T>::less |
| 56 | { |
| 57 | const Self & parent; |
| 58 | int nan_direction_hint; |
| 59 | less(const Self & parent_, int nan_direction_hint_) : parent(parent_), nan_direction_hint(nan_direction_hint_) {} |
| 60 | bool operator()(size_t lhs, size_t rhs) const { return CompareHelper<T>::less(parent.data[lhs], parent.data[rhs], nan_direction_hint); } |
| 61 | }; |
| 62 | |
| 63 | template <typename T> |
| 64 | struct ColumnVector<T>::greater |
| 65 | { |
| 66 | const Self & parent; |
| 67 | int nan_direction_hint; |
| 68 | greater(const Self & parent_, int nan_direction_hint_) : parent(parent_), nan_direction_hint(nan_direction_hint_) {} |
| 69 | bool operator()(size_t lhs, size_t rhs) const { return CompareHelper<T>::greater(parent.data[lhs], parent.data[rhs], nan_direction_hint); } |
| 70 | }; |
| 71 | |
| 72 | |
| 73 | namespace |
| 74 | { |
| 75 | template <typename T> |
| 76 | struct ValueWithIndex |
| 77 | { |
| 78 | T value; |
| 79 | UInt32 index; |
| 80 | }; |
| 81 | |
| 82 | template <typename T> |
| 83 | struct RadixSortTraits : RadixSortNumTraits<T> |
| 84 | { |
| 85 | using Element = ValueWithIndex<T>; |
| 86 | static T & (Element & elem) { return elem.value; } |
| 87 | }; |
| 88 | } |
| 89 | |
| 90 | template <typename T> |
| 91 | void ColumnVector<T>::getPermutation(bool reverse, size_t limit, int nan_direction_hint, IColumn::Permutation & res) const |
| 92 | { |
| 93 | size_t s = data.size(); |
| 94 | res.resize(s); |
| 95 | |
| 96 | if (s == 0) |
| 97 | return; |
| 98 | |
| 99 | if (limit >= s) |
| 100 | limit = 0; |
| 101 | |
| 102 | if (limit) |
| 103 | { |
| 104 | for (size_t i = 0; i < s; ++i) |
| 105 | res[i] = i; |
| 106 | |
| 107 | if (reverse) |
| 108 | std::partial_sort(res.begin(), res.begin() + limit, res.end(), greater(*this, nan_direction_hint)); |
| 109 | else |
| 110 | std::partial_sort(res.begin(), res.begin() + limit, res.end(), less(*this, nan_direction_hint)); |
| 111 | } |
| 112 | else |
| 113 | { |
| 114 | /// A case for radix sort |
| 115 | if constexpr (is_arithmetic_v<T> && !std::is_same_v<T, UInt128>) |
| 116 | { |
| 117 | /// Thresholds on size. Lower threshold is arbitrary. Upper threshold is chosen by the type for histogram counters. |
| 118 | if (s >= 256 && s <= std::numeric_limits<UInt32>::max()) |
| 119 | { |
| 120 | PaddedPODArray<ValueWithIndex<T>> pairs(s); |
| 121 | for (UInt32 i = 0; i < s; ++i) |
| 122 | pairs[i] = {data[i], i}; |
| 123 | |
| 124 | RadixSort<RadixSortTraits<T>>::executeLSD(pairs.data(), s); |
| 125 | |
| 126 | /// Radix sort treats all NaNs to be greater than all numbers. |
| 127 | /// If the user needs the opposite, we must move them accordingly. |
| 128 | size_t nans_to_move = 0; |
| 129 | if (std::is_floating_point_v<T> && nan_direction_hint < 0) |
| 130 | { |
| 131 | for (ssize_t i = s - 1; i >= 0; --i) |
| 132 | { |
| 133 | if (isNaN(pairs[i].value)) |
| 134 | ++nans_to_move; |
| 135 | else |
| 136 | break; |
| 137 | } |
| 138 | } |
| 139 | |
| 140 | if (reverse) |
| 141 | { |
| 142 | if (nans_to_move) |
| 143 | { |
| 144 | for (size_t i = 0; i < s - nans_to_move; ++i) |
| 145 | res[i] = pairs[s - nans_to_move - 1 - i].index; |
| 146 | for (size_t i = s - nans_to_move; i < s; ++i) |
| 147 | res[i] = pairs[s - 1 - (i - (s - nans_to_move))].index; |
| 148 | } |
| 149 | else |
| 150 | { |
| 151 | for (size_t i = 0; i < s; ++i) |
| 152 | res[s - 1 - i] = pairs[i].index; |
| 153 | } |
| 154 | } |
| 155 | else |
| 156 | { |
| 157 | if (nans_to_move) |
| 158 | { |
| 159 | for (size_t i = 0; i < nans_to_move; ++i) |
| 160 | res[i] = pairs[i + s - nans_to_move].index; |
| 161 | for (size_t i = nans_to_move; i < s; ++i) |
| 162 | res[i] = pairs[i - nans_to_move].index; |
| 163 | } |
| 164 | else |
| 165 | { |
| 166 | for (size_t i = 0; i < s; ++i) |
| 167 | res[i] = pairs[i].index; |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | return; |
| 172 | } |
| 173 | } |
| 174 | |
| 175 | /// Default sorting algorithm. |
| 176 | for (size_t i = 0; i < s; ++i) |
| 177 | res[i] = i; |
| 178 | |
| 179 | if (reverse) |
| 180 | pdqsort(res.begin(), res.end(), greater(*this, nan_direction_hint)); |
| 181 | else |
| 182 | pdqsort(res.begin(), res.end(), less(*this, nan_direction_hint)); |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | |
| 187 | template <typename T> |
| 188 | const char * ColumnVector<T>::getFamilyName() const |
| 189 | { |
| 190 | return TypeName<T>::get(); |
| 191 | } |
| 192 | |
| 193 | template <typename T> |
| 194 | MutableColumnPtr ColumnVector<T>::cloneResized(size_t size) const |
| 195 | { |
| 196 | auto res = this->create(); |
| 197 | |
| 198 | if (size > 0) |
| 199 | { |
| 200 | auto & new_col = static_cast<Self &>(*res); |
| 201 | new_col.data.resize(size); |
| 202 | |
| 203 | size_t count = std::min(this->size(), size); |
| 204 | memcpy(new_col.data.data(), data.data(), count * sizeof(data[0])); |
| 205 | |
| 206 | if (size > count) |
| 207 | memset(static_cast<void *>(&new_col.data[count]), static_cast<int>(ValueType()), (size - count) * sizeof(ValueType)); |
| 208 | } |
| 209 | |
| 210 | return res; |
| 211 | } |
| 212 | |
| 213 | template <typename T> |
| 214 | UInt64 ColumnVector<T>::get64(size_t n) const |
| 215 | { |
| 216 | return ext::bit_cast<UInt64>(data[n]); |
| 217 | } |
| 218 | |
| 219 | template <typename T> |
| 220 | Float64 ColumnVector<T>::getFloat64(size_t n) const |
| 221 | { |
| 222 | return static_cast<Float64>(data[n]); |
| 223 | } |
| 224 | |
| 225 | template <typename T> |
| 226 | Float32 ColumnVector<T>::getFloat32(size_t n) const |
| 227 | { |
| 228 | return static_cast<Float32>(data[n]); |
| 229 | } |
| 230 | |
| 231 | template <typename T> |
| 232 | void ColumnVector<T>::insertRangeFrom(const IColumn & src, size_t start, size_t length) |
| 233 | { |
| 234 | const ColumnVector & src_vec = assert_cast<const ColumnVector &>(src); |
| 235 | |
| 236 | if (start + length > src_vec.data.size()) |
| 237 | throw Exception("Parameters start = " |
| 238 | + toString(start) + ", length = " |
| 239 | + toString(length) + " are out of bound in ColumnVector<T>::insertRangeFrom method" |
| 240 | " (data.size() = " + toString(src_vec.data.size()) + ")." , |
| 241 | ErrorCodes::PARAMETER_OUT_OF_BOUND); |
| 242 | |
| 243 | size_t old_size = data.size(); |
| 244 | data.resize(old_size + length); |
| 245 | memcpy(data.data() + old_size, &src_vec.data[start], length * sizeof(data[0])); |
| 246 | } |
| 247 | |
| 248 | template <typename T> |
| 249 | ColumnPtr ColumnVector<T>::filter(const IColumn::Filter & filt, ssize_t result_size_hint) const |
| 250 | { |
| 251 | size_t size = data.size(); |
| 252 | if (size != filt.size()) |
| 253 | throw Exception("Size of filter doesn't match size of column." , ErrorCodes::SIZES_OF_COLUMNS_DOESNT_MATCH); |
| 254 | |
| 255 | auto res = this->create(); |
| 256 | Container & res_data = res->getData(); |
| 257 | |
| 258 | if (result_size_hint) |
| 259 | res_data.reserve(result_size_hint > 0 ? result_size_hint : size); |
| 260 | |
| 261 | const UInt8 * filt_pos = filt.data(); |
| 262 | const UInt8 * filt_end = filt_pos + size; |
| 263 | const T * data_pos = data.data(); |
| 264 | |
| 265 | #ifdef __SSE2__ |
| 266 | /** A slightly more optimized version. |
| 267 | * Based on the assumption that often pieces of consecutive values |
| 268 | * completely pass or do not pass the filter. |
| 269 | * Therefore, we will optimistically check the parts of `SIMD_BYTES` values. |
| 270 | */ |
| 271 | |
| 272 | static constexpr size_t SIMD_BYTES = 16; |
| 273 | const __m128i zero16 = _mm_setzero_si128(); |
| 274 | const UInt8 * filt_end_sse = filt_pos + size / SIMD_BYTES * SIMD_BYTES; |
| 275 | |
| 276 | while (filt_pos < filt_end_sse) |
| 277 | { |
| 278 | int mask = _mm_movemask_epi8(_mm_cmpgt_epi8(_mm_loadu_si128(reinterpret_cast<const __m128i *>(filt_pos)), zero16)); |
| 279 | |
| 280 | if (0 == mask) |
| 281 | { |
| 282 | /// Nothing is inserted. |
| 283 | } |
| 284 | else if (0xFFFF == mask) |
| 285 | { |
| 286 | res_data.insert(data_pos, data_pos + SIMD_BYTES); |
| 287 | } |
| 288 | else |
| 289 | { |
| 290 | for (size_t i = 0; i < SIMD_BYTES; ++i) |
| 291 | if (filt_pos[i]) |
| 292 | res_data.push_back(data_pos[i]); |
| 293 | } |
| 294 | |
| 295 | filt_pos += SIMD_BYTES; |
| 296 | data_pos += SIMD_BYTES; |
| 297 | } |
| 298 | #endif |
| 299 | |
| 300 | while (filt_pos < filt_end) |
| 301 | { |
| 302 | if (*filt_pos) |
| 303 | res_data.push_back(*data_pos); |
| 304 | |
| 305 | ++filt_pos; |
| 306 | ++data_pos; |
| 307 | } |
| 308 | |
| 309 | return res; |
| 310 | } |
| 311 | |
| 312 | template <typename T> |
| 313 | ColumnPtr ColumnVector<T>::permute(const IColumn::Permutation & perm, size_t limit) const |
| 314 | { |
| 315 | size_t size = data.size(); |
| 316 | |
| 317 | if (limit == 0) |
| 318 | limit = size; |
| 319 | else |
| 320 | limit = std::min(size, limit); |
| 321 | |
| 322 | if (perm.size() < limit) |
| 323 | throw Exception("Size of permutation is less than required." , ErrorCodes::SIZES_OF_COLUMNS_DOESNT_MATCH); |
| 324 | |
| 325 | auto res = this->create(limit); |
| 326 | typename Self::Container & res_data = res->getData(); |
| 327 | for (size_t i = 0; i < limit; ++i) |
| 328 | res_data[i] = data[perm[i]]; |
| 329 | |
| 330 | return res; |
| 331 | } |
| 332 | |
| 333 | template <typename T> |
| 334 | ColumnPtr ColumnVector<T>::index(const IColumn & indexes, size_t limit) const |
| 335 | { |
| 336 | return selectIndexImpl(*this, indexes, limit); |
| 337 | } |
| 338 | |
| 339 | template <typename T> |
| 340 | ColumnPtr ColumnVector<T>::replicate(const IColumn::Offsets & offsets) const |
| 341 | { |
| 342 | size_t size = data.size(); |
| 343 | if (size != offsets.size()) |
| 344 | throw Exception("Size of offsets doesn't match size of column." , ErrorCodes::SIZES_OF_COLUMNS_DOESNT_MATCH); |
| 345 | |
| 346 | if (0 == size) |
| 347 | return this->create(); |
| 348 | |
| 349 | auto res = this->create(); |
| 350 | typename Self::Container & res_data = res->getData(); |
| 351 | res_data.reserve(offsets.back()); |
| 352 | |
| 353 | IColumn::Offset prev_offset = 0; |
| 354 | for (size_t i = 0; i < size; ++i) |
| 355 | { |
| 356 | size_t size_to_replicate = offsets[i] - prev_offset; |
| 357 | prev_offset = offsets[i]; |
| 358 | |
| 359 | for (size_t j = 0; j < size_to_replicate; ++j) |
| 360 | res_data.push_back(data[i]); |
| 361 | } |
| 362 | |
| 363 | return res; |
| 364 | } |
| 365 | |
| 366 | template <typename T> |
| 367 | void ColumnVector<T>::gather(ColumnGathererStream & gatherer) |
| 368 | { |
| 369 | gatherer.gather(*this); |
| 370 | } |
| 371 | |
| 372 | template <typename T> |
| 373 | void ColumnVector<T>::getExtremes(Field & min, Field & max) const |
| 374 | { |
| 375 | size_t size = data.size(); |
| 376 | |
| 377 | if (size == 0) |
| 378 | { |
| 379 | min = T(0); |
| 380 | max = T(0); |
| 381 | return; |
| 382 | } |
| 383 | |
| 384 | bool has_value = false; |
| 385 | |
| 386 | /** Skip all NaNs in extremes calculation. |
| 387 | * If all values are NaNs, then return NaN. |
| 388 | * NOTE: There exist many different NaNs. |
| 389 | * Different NaN could be returned: not bit-exact value as one of NaNs from column. |
| 390 | */ |
| 391 | |
| 392 | T cur_min = NaNOrZero<T>(); |
| 393 | T cur_max = NaNOrZero<T>(); |
| 394 | |
| 395 | for (const T x : data) |
| 396 | { |
| 397 | if (isNaN(x)) |
| 398 | continue; |
| 399 | |
| 400 | if (!has_value) |
| 401 | { |
| 402 | cur_min = x; |
| 403 | cur_max = x; |
| 404 | has_value = true; |
| 405 | continue; |
| 406 | } |
| 407 | |
| 408 | if (x < cur_min) |
| 409 | cur_min = x; |
| 410 | else if (x > cur_max) |
| 411 | cur_max = x; |
| 412 | } |
| 413 | |
| 414 | min = NearestFieldType<T>(cur_min); |
| 415 | max = NearestFieldType<T>(cur_max); |
| 416 | } |
| 417 | |
| 418 | /// Explicit template instantiations - to avoid code bloat in headers. |
| 419 | template class ColumnVector<UInt8>; |
| 420 | template class ColumnVector<UInt16>; |
| 421 | template class ColumnVector<UInt32>; |
| 422 | template class ColumnVector<UInt64>; |
| 423 | template class ColumnVector<UInt128>; |
| 424 | template class ColumnVector<Int8>; |
| 425 | template class ColumnVector<Int16>; |
| 426 | template class ColumnVector<Int32>; |
| 427 | template class ColumnVector<Int64>; |
| 428 | template class ColumnVector<Int128>; |
| 429 | template class ColumnVector<Float32>; |
| 430 | template class ColumnVector<Float64>; |
| 431 | } |
| 432 | |